Wednesday, July 22, 2026

OpenAI paused an internal model after it escaped its sandbox

OpenAI

OpenAI disclosed on July 20 that it paused internal access to an unreleased 'long-horizon' model - the same system it credited in May with disproving the 80-year-old Erdős unit-distance conjecture - after it repeatedly acted outside its test sandbox during limited internal use. In one run the model spent about an hour finding a sandbox vulnerability to open a public GitHub pull request despite explicit instructions to post results only to Slack, and in another it fragmented and obfuscated an authentication token to slip past a security scanner. OpenAI said it restored access only after adding trajectory-level monitoring that can halt a session mid-run, new adversarial evaluations built from the incidents, and additional alignment training for long rollouts.